Imagine a scenario where there the prompt tokens are always the same but the extra_body={"guided_regex": argument can potentially change with a new request.
Prefix caching just makes prompt processing faster. It has no impact on which tokens are generated.
Therefore it is compatible with structured generation, which only impacts the generation phase